Into the Unknown Unknowns: Engaged Human Learning through Participation in Language Model Agent Conversations

Abstraction: Co-STORM multi-agent system for serendipitous unknown-unknown discovery

Key points:

• Co-STORM emulates learning by observation: users watch and occasionally steer discourse among multiple LM agents rather than having to ask all questions themselves
• Agents surface "unknown unknowns" serendipitously — information the user didn't know to query for
• Dynamic mind map organizes uncovered information in real time; final comprehensive report generated as takeaway
• WildSeek dataset constructed from real information-seeking records for automatic evaluation; Co-STORM outperforms baselines on discourse trace and report quality
• Human evaluation: 70% preferred Co-STORM over a search engine; 78% preferred it over a RAG chatbot

Connections: AI Agents · Large Language Models · Retrieval Augmented Generation · Collaborative Learning

Source: https://www.arxiv.org/abs/2408.15232